Background: Late gadolinium enhancement (LGE) in cardiovascular magnetic resonance (CMR) is critical for accurate tissue characterization, particularly for identifying areas of scar or infarction. Optimal inversion time (TI) selection, where the myocardium is adequately suppressed and contrast maximized, is traditionally performed manually by technologists using the TI scout sequence, which can introduce variability. This study aims to evaluate the potential of an automated method to assist technologists from a high-volume center in achieving consistent and precise TI selection.
METHODS AND RESULTS: We retrospectively analyzed 85 CMR exams performed at 3T Signa Premier (GE HealthCare), including healthy individuals, ischemic cardiomyopathy (ICMP), non-ischemic cardiomyopathy (NICMP), and amyloidosis cases. A deep learning-based segmentation model was used to delineate the blood pool and myocardium in the TI scout sequence images. For each frame, the mean signal intensities of the blood pool and myocardium were calculated and divided to generate a contrast ratio. The frame with the highest contrast ratio was identified as the optimal TI, which was then compared to the TI selected by an expert technologist from a high volume center. Additionally, 10 scans were evaluated with both the technologist-selected TI and a -30 ms offset (determined by the primary analysis) to assess algorithm performance.
The automated TI selection method showed a mean difference of 27.12 ms earlier and a mean absolute error (MAE) of 42.17 ms compared to technologist-selected TIs, with an accuracy within 60 ms in 87.06% of cases. In 90.59% of cases, the algorithm-selected TI was at or earlier than the technologist's choice (Figure 1). In 9 out of 10 cases, expert readers preferred the -30 ms offset, noting superior contrast. Figure 2 shows examples of normal myocardium and amyloidosis. In amyloidosis, the myocardium peaks before the blood pool (reverse kinetics), contrary to the usual pattern in non-amyloidosis cases. The algorithm successfully identified this characteristic, demonstrating its ability to detect pathophysiological differences.
Conclusion: The automated segmentation and contrast optimization algorithm showed accuracy comparable to experienced technologists and, in some cases, provided better contrast. It also identified reverse kinetics in amyloidosis cases, which may have clinical significance. Further validation could establish this tool as a way to improve consistency and reduce operator dependency in LGE imaging.
